medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Quantifying the Benefits of Targeting for Pandemic Response Sergio Cameloa, Dragos F. Ciocanb, Dan A. Iancub,c, Xavier S. Warnesc, and Spyros I. Zoumpoulisd aInstitute for Computational and Mathematical Engineering, Stanford University bTechnology and Operations Management, INSEAD cOperations, Information & Technology, Graduate School of Business, Stanford University dDecision Sciences, INSEAD
ABSTRACT
To respond to pandemics such as COVID-19, policy makers have relied on interventions that target specific age groups or activities. Such targeting is potentially contentious, so rigorously quantifying its benefits and downsides is critical for designing effective and equitable pandemic control policies. We propose a flexible modeling framework and algorithms to compute optimally targeted interventions that coordinate across two dimensions of heterogeneity: age of different groups and the specific activities that individuals engage in during the course of a day. We showcase a complete implementation in a case study focused on the COVID-19 epidemic in the Île-de-France region of France, based on hospitalization, community mobility, social contacts and economic data. We find that optimized dual-targeted policies generate substantial complementarities that lead to Pareto improvements, reducing the number of deaths and the economic losses overall and reducing the time in confinement for each age group, compared to less targeted interventions. These policies have a simple and explainable structure. Since dual-targeted policies could lead to increased discrepancies in the confinements faced by distinct groups, we also quantify the impact of requirements that explicitly limit such disparities, and find that satisfactory trade-offs may be achievable through limited targeting.
Keywords: Pandemic management, Confinement, Targeted interventions, Optimization, COVID-19
Introduction The COVID-19 pandemic has forced policy makers worldwide to rely on a range of large-scale population confinement measures in an effort to contain the disease spread. In determining these measures, a key recognition has been that substantial differences exist in the health and economic impact produced by different individuals engaged in distinct activities. Targeting confinements to account for such heterogeneity could be an important lever to mitigate a pandemic’s impact, but could also lead to potentially contentious and discriminatory measures. This work is aimed at developing a rigorous framework to quantify the benefits and downsides of such targeted interventions, and applying it to the COVID-19 pandemic as a real-world case study. One real-world contentious example of targeting has been to differentiate confinements based on age groups, e.g., sheltering older individuals who might face higher health risks if infected, or restricting younger groups who might create higher infection risks. Such measures, focusing on confinements or other interventions, and targeting age or other population characteristics, have been studied in the literature (1; 2; 3; 4; 5; 6; 7; 8; 9; 10) and implemented in several settings – e.g., with stricter confinements applied to older groups in Finland (11), Ireland (12), Israel (13) and Moscow (14), or curfews applied to children and youth in Bosnia and Herzegovina (15) and Turkey (16) – but some of the measures were eventually deemed unconstitutional and overturned (13; 15). A different example of targeting has stemmed from the recognition that different activities (more specifically, population interactions in locations of certain activities), such as work, schooling, transport, leisure, result in significantly different patterns of social contacts and new infections. This has been shown to be critical when modeling pandemic spread (17; 18; 19) and has been recognized in numerous implementations that differentially confine various activities (e.g., closures of schools, workplaces, recreation venues, etc.), and even some that differentiate based on both age groups and activities (e.g., dedicated hours when only the senior population was allowed to shop at supermarkets (20)). As these examples suggest, targeted confinements have merits but also pose potentially significant downsides. On the one hand, targeting can generate improvements in both health and economic outcomes, giving policy makers an improved lever when navigating difficult trade-offs. Additionally, explicitly considering multiple dimensions of targeting simultaneously – activities and age groups – could overturn some of the prevailing insight that specific age groups should uniformly face stricter confinements. However, such granular policies are more difficult to implement, and could lead to discriminatory and potentially unfair measures. Given that some amount of targeting of activities and age groups is already in place in existing real-world
NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice. medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
policy implementations, it seems critical to quantify the relative merits of a policy that (i) targets both age-based population groups and activities, and (ii) identifies optimal interventions. We propose a rigorous modeling framework and develop a set of associated algorithms that compute optimally targeted interventions that target both age groups and activities. The framework provides quantifiable answers to the following natural research questions: How large are the health and economic benefits of dual-targeted confinements? Would dual targeting lead to significant synergies, and why? Could dual targeting reduce time in confinement for every age group? What is the relationship between the effectiveness and the level of targeting allowed across distinct groups? We showcase a proof of concept for our framework through a case study calibrated on Île-de-France data – a region of France encompassing Paris with a population of approximately 12 million. The implementation, publicly available at http://insead.arnia.ro, is flexible and portable to other geographies.
Methods Our framework relies on a flexible model that captures several important real-world considerations. We extend a version of the discretized SEIR (Susceptible-Exposed-Infectious-Recovered) epidemiological model (18; 21; 22) with multiple population groups that interact with each other (SI §2). We augment the model with controls that target based on (i) age groups, and (ii) types of activities that individuals engage in. Different policy interventions can be embedded as controls: we focus on time-dependent, targeted confinements, but extend the model to mass testing in the SI; vaccinations can also be accommodated. Interventions modulate the rate of social contacts and the economic value generated, and the objective of the control problem is to minimize a combination of health and economic losses caused by deaths, illness, and activity restrictions. The model captures important resource constraints (such as hospital and ICU), and allows explicitly controlling the amount of targeting through “limited disparity” constraints that limit the difference in the extent of confinement imposed on distinct age groups. Epidemiological Model and Controls. We segment the population by age into nine groups g G ; the youngest group g 2 captures individuals with age 0-9 and the oldest those aged 80 or above. For each g, the compartmental model includes states for susceptible, exposed, infectious, quarantined infectious, recovered, and deceased individuals. We also reserve separate states for individuals who are hospitalized due to being infected, in either general hospital wards or in intensive care units (ICU). We use T to denote the time horizon of the control problem, and X t to denote the entire vector of epidemiological states at time 0 t T. Individuals interact in activities belonging to the set A = work,transport,leisure,school,home,other . These interac- { } tions generate social contacts which drive the rate of new infections. We control the SEIR dynamics by adjusting the confinement intensity in each group-activity pair over time: we let a `g(t) [0,1] denote the activity level allowed for group g and activity a at time t, expressed as a fraction of the activity level 2 a under normal course of life (no confinement). We denote `g(t)=[`g(t)]a A , ut =[`g(t)]g G and ut:t =[ut ,...,ut ]. 2 2 0 0 We propose a parametric model to map activity levels to social contacts. We use cg,h(`g,`h) to denote the mean number of total daily contacts between an individual in group g and individuals in group h across all activities when their activity levels are `g,`h, respectively. Varying the activity levels changes the social contacts according to
a a a1 a a2 cg,h(`g,`h)= Â Cg,h (`g) (`h) , (1) a A · · 2 a where Cg,h denote the mean number of daily contacts in activity a under normal course (i.e., without confinement), and a a1,a2 R are parameters. We retrieve values for C from the data tool (23), which is based on the French social contact 2 g,h survey data in (24), and we estimate a1,a2 from health outcome data (25) and Google mobility data (26). When the number of patients requiring hospitalization or ICU treatment exceeds the respective capacity of available beds, we assume that patients are turned away from each age group according to a proportional rule.1 Objective. Our objective captures two criteria. The first quantifies the total deaths directly attributable to the pandemic, which we denote by Total Deaths(u0:T 1) to reflect the dependency on the specific confinement policy u0:T 1 followed. The second captures the economic losses due to the pandemic, denoted by Economic Loss(u0:T 1) and stemming from work restrictions of activity (SI §2). For example, lowering `30 39 y.o.(t) causes lost wages in g = 30 39 y.o., while lowering `school (t) causes lost schooling costs for g = 10 19 y.o. 10 19 y.o. To allow policy makers to weight the importance of the two criteria, we associate a cost c to each death, which we express in multiples of French GDP per capita. Our framework can capture a multitude of policy preferences by considering a wide range of c values, from completely prioritizing economic losses (c = 0) to completely prioritizing deaths (c •). ! 1Our framework allows implementing any turn-away rule or even optimizing over these decisions.
2/9 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Optimization Problem. We seek control policies u0:T 1 that minimize the sum of economic and mortality costs. More formally, we solve:
minu0:T 1 Economic Loss(u0:T 1)+c Total Deaths(u0:T 1) (2) · subject to constraints⇥ that (i) the state trajectory follows the SEIR dynamics,⇤ and (ii) the controls and states do not exceed the available capacities of ICU beds and hospital beds. Re-Optimization with Linearized Dynamics — ROLD. Minimizing (2) is challenging to solve to optimality even for small time horizons due to the non-linear and non-convex nature of the SEIR dynamics. To overcome this, we design a model predictive control (27) approximate algorithm that remains tractable for realistic time horizons and problem sizes. The algorithm proceeds sequentially over the time horizon. When the SEIR state is X t at time t, it builds an approximation of the SEIR dynamics from t to T that is linear in the controls ut:T 1. This linearity allows us to find the best control ut⇤:T 1 with respect to these linearized dynamics. The algorithm uses ut⇤ as the intervention in the real SEIR system to advance to state X t+1, and iterates by updating the linearization around the new state. Re-solving at each time t is aimed to prevent the linearization errors from growing too large (details in SI §3). Parametrization and Model Calibration. We adopt the confidence regions for SEIR parameters reported in (22) for the Île-de-France region, which we complement with Google mobility data to approximate the mean effective lockdowns for activities during the horizon of interest. We estimate values of all parameters by comparing the simulated SEIR model against real data on hospital and ICU utilization and deaths. We calibrate our economic model using data from France (and where available Île-de-France) on full time equivalent wages and employment rates, and sentiment surveys on business activity levels during confinement. Experimental Setup. We use a horizon of T = 90 in the experiments reported in the main paper, and allow up to T = 360 days in additional experiments reported in SI. We allow the confinement decisions to change every two weeks. We use a capacity of 2900 beds for ICU in Île-de-France, and infinite beds for general hospital wards. We consider several ROLD policies that differ in the level of targeting allowed, which we compare over a wide range of values for c, from 0 to 1000 the annual GDP per ⇥ capita in France2. In SI §7 we also discuss the additional benefits of targeted administration of viral tests across age groups. We provide all the details for the ROLD heuristic, calibration, parameter specification, and experimental setup in the SI, where we also report results from sensitivity and robustness analyses on the fitted parameters.
Results We use this framework to address our main research questions. How large are the gains from dual targeting? To isolate the benefits of each type of targeting, we compare four versions of ROLD that differ in the level of targeting allowed: no targeting whatsoever (“NO-TARGET”), targeting age groups only (“AGE”), activities only (“ACT”), or both (“AGE-ACT”, or simply “ROLD” when no confusion can arise). Figure 1a records each policy’s performance in several problem instances parameterized by the cost of death c. A striking feature is that each of the targeted policies actually Pareto-dominates the NO-TARGET policy, and the improvements are significant: relative to NO-TARGET and for same number of deaths, economic losses are reduced by EUR 0-2.9B (0%-35.9%) in AGE, by EUR 0.4B-2.1B (4.5%-49.8%) in ACT, and by EUR 3.3B-5.3B (35.7%-80.0%) in AGE-ACT. This Pareto-dominance is unexpected, since it is not a property that we explicitly ask for in our optimization procedure, and it underlines that any form of targeting can lead to significant improvements in terms of both health and economic outcomes. When comparing the different types of targeting, neither AGE nor ACT Pareto-dominate each other, and neither policy dominates in terms of the total loss objective (SI Figure S3). In contrast and crucially, AGE-ACT Pareto-dominates all other policies, and moreover leads to super-additive improvements in almost all cases: for the same number of deaths, AGE- ACT reduces economic losses by more than AGE and ACT added together (SI Figure S4). This suggests that substantial complementarities may be unlocked through targeting both age groups and activities, which may not be available under less granular targeting. To confirm the significance of these gains, we also compare ROLD AGE-ACT with various practical benchmark policies in Figure 1b. Benchmarks ICU-t and Hybrid-t AND/Hybrid-t OR mimic implementations in the U.S. Austin area (28) and, respectively, France (29). These policies switch between a stricter and a relaxed confinement level based on conditions related to hospital occupancy and the rate of new infections (SI §5). We also consider two extreme benchmarks corresponding to enforcing “full confinement” (FC) or remaining “fully open” (FO); these can be expected to perform well when completely prioritizing one of the two metrics of interest, with FC minimizing the number of deaths and FO ensuring low economic losses.
2We quantify the cost of death c as a multiple of the annual GDP per capita in France, and use the shorthand notation n to denote a value of n times this ⇥ annual GDP per capita.
3/9 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
ROLD Pareto-dominates all these benchmarks, decreasing economic losses by EUR 5.3B-16.9B (71.0%-82.6%) relative to Hybrid-t AND, by EUR 7.1B-11.6B (62.2%-82.8%) relative to Hybrid-t OR, and by EUR 5.4B-11.6B (62.2%-78.0%) relative to ICU-t for the same number of deaths. Additionally, ROLD meets or exceeds the performance of the two extreme policies: for a sufficiently large c, ROLD exactly recovers the FC policy, resulting in 890 deaths and economic losses of EUR 27.6B; for a sufficiently low c, ROLD actually Pareto-dominates the FO policy, reducing the number of deaths by 16,688 (76.7%) and reducing economic losses by EUR 1.6B (65.3%). The latter result, which may seem surprising, is driven by the natural premise captured in our model that deaths and illness generate economic loss because of lost productivity; thus, a smart sequence of confinement decisions can actually improve the economic loss relative to FO. Among all the policies we tested, ROLD AGE-ACT was the only one capable of Pareto-dominating the FO benchmark, confirming that dual targeting is critical and powerful. The Pareto-dominance of ROLD AGE-ACT implies its dominance in terms of the total loss objective (SI Figure S3). These results are robust under more problem instances (SI Figure S5).
(a) Comparison between four ROLD policies with different levels of (b) Comparison between ROLD with dual targeting and benchmark targeting policies
Figure 1. The total number of deaths and the economic losses generated by ROLD policies with different levels of targeting and by the benchmark policies. Panel (a) compares the four versions of ROLD that differ in the level of targeting allowed. Panel (b) compares the ROLD policy that targets age groups and activities with the benchmark policies. Each marker corresponds to a different problem instance parametrized by the cost of death c. We include 128 distinct values of c from 0 to 990 , and panel (b) also includes a very large value (c = 1016 ). ⇥ ⇥ How do gains arise from dual targeting? We examine the structure of the optimal ROLD AGE-ACT confinement decisions. We focus our discussion on the value c = 50 , which is in the mid-range of estimates used in the economics literature on ⇥ COVID-19 (30) and is representative of the behavior we observe across all experiments (SI §7). Figure 2a visualizes the optimized confinement policy. Generally, the ROLD policy maintains high activity levels for those groups with a high ratio of marginal economic value to total social contacts in the activity, i.e., a high a dvg(ut )/d`g(t) “econ-to-contacts-ratio” := a , Âh G Ch,g 2 where vg(ut ) is the economic value created by an individual in age group g when activity levels are ut . For example, in work, ROLD completely opens up the 40-69 y.o. groups, while confining the 20-39 y.o. groups during the first two weeks and the 10-19 y.o. groups for the first ten weeks. This is explainable since the 40-69 y.o. age groups produce the highest econ-to-contacts-ratio in work, while the younger groups have progressively lower ratios. Similarly, ROLD prioritizes activity in transport, then other, then leisure, in accordance with the relative econ-to-contacts-ratio of these activities. To confirm the robustness of this insight, we also conduct a more thorough study where we compute optimal ROLD policies for several problem instances, and then train regression decision trees to predict the optimal ROLD activity levels as a function of several features (SI §7). The results are captured in Figure 2b and SI Figure S10. These simple trees can predict the optimal ROLD activity levels quite well (with root MSE values in the range 0.10-0.22), and they confirm our core insight that the econ-to-contacts ratio is the most salient feature when targeting confinements, as it is used as a split variable in the root node of each tree, with higher ratios leading to higher activity levels in all activities considered. To understand how complementarities arise in this context, note that the ability to separately target age groups and activities allows the ROLD policy to fully exploit the fact that distinct age groups may be responsible for the largest econ-to-contacts
4/9 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
ratio in different activities. As an example, the 20-69 y.o. groups have the highest ratio in work, whereas the 0-19 y.o. and 70+ y.o. groups have the highest ratio in leisure. ROLD coordinates confinements to account for this: groups 20-69 y.o. remain more open in work but face confinement in leisure for the first ten weeks, whereas the 10-19 y.o. group is confined in work for a long period while remaining open in leisure, as remain the 70+ y.o. groups. These complementary confinement schedules allow ROLD to reduce both the number of deaths and economic losses, with the added benefit that no age group is completely confined.
samples = 9,240 activity level = 0.00% samples = 1,575 activity level = 8.25% Decision tree for "work" RMSE: 0.18439 econ-to-contacts-ratio <= 60.6 time <= 63.0 samples = 630 samples = 11,445 samples = 2,205 True activity level = 100.00% activity level = 6.64% activity level = 34.46% econ-to-contacts-ratio <= 176.9 samples = 27,720 samples = 2,010 activity level = 55.85% time <= 21.0 econ-to-contacts-ratio <= 765.4 activity level = 43.45% False samples = 16,275 samples = 4,650 activity level = 90.45% activity level = 69.48% samples = 2,640 activity level = 89.31%
econ-to-contacts-ratio <= 305.9 samples = 625 samples = 11,625 activity level = 81.60% activity level = 98.83%
samples = 11,000 activity level = 99.81%
samples = 7,875 activity level = 0.00% samples = 1,520 activity level = 5.71% Decision tree for "leisure" RMSE: 0.21767 econ-to-contacts-ratio <= 51.4 time <= 49.0 samples = 1,140 samples = 10,535 samples = 2,660 True activity level = 86.79% activity level = 10.22% activity level = 40.46% econ-to-contacts-ratio <= 113.3 samples = 27,720 samples = 1,640 activity level = 56.67% time <= 21.0 econ-to-contacts-ratio <= 226.5 activity level = 23.87% False samples = 17,185 samples = 4,910 activity level = 85.14% activity level = 60.86% samples = 3,270 activity level = 79.41%
econ-to-contacts-ratio <= 213.8 samples = 3,475 samples = 12,275 activity level = 82.51% activity level = 94.86%
samples = 8,800 activity level = 99.74%
Decision tree for "other" RMSE: 0.13489 samples = 2,250 activity level = 89.86% True samples = 7,875 activity level = 0.00% econ-to-contacts-ratio <= 158.4 χ <= 20.5 samples = 27,720 samples = 2,835 samples = 585 activity level = 68.25% time <= 7.0 activity level = 77.33% activity level = 29.12% False samples = 19,845 activity level = 95.33% time <= 21.0 samples = 2,835 samples = 17,010 activity level = 91.69% activity level = 98.33%
samples = 14,175 activity level = 99.66% (a) Optimized ROLD confinement policy and its estimated impact (b) Decision trees for work, leisure and other activities
Figure 2. The optimized ROLD AGE-ACT policies for problem instances with a 90-day optimization horizon starting on October 21, 2020 (see SI for optimized ROLD policies with a 360-day optimization horizon). Figure (a) corresponds to a problem instance where the cost of death c is 50 . The seven panels depict the time evolution for the occupation of hospital ⇥ and ICU beds (top panel), the number of actively infectious individuals and the cumulative number of deceased individuals in the population (panel 2), and the confinement policy imposed by ROLD in each age group and activity (panels 3-7). In panels 3-7, the values correspond to the activity levels allowed for the respective age group, and are color-coded so that darker shades capture a stricter confinement. Figure (b) depicts decision trees approximating the optimized ROLD confinement decisions for work, leisure and other (trained with 27,720 samples), with a horizon of T = 90 days. Each node in the tree records several pieces of information: a logical condition based on which all the training samples in the node are split, with the upper sub-tree corresponding to the logical condition being true (e.g., “econ-to-contacts-ratio 176.9” for the root node in the work tree), the number of training samples falling in the node (“samples”), and the average activity level for all the samples in the node. The nodes are color-coded based on the activity level, with darker colors corresponding to stricter confinement.
Can dual targeting reduce time in confinement for each age group? We calculate the fraction of time spent by each age group in confinement under each ROLD policy, averaged over the activities relevant to that age group (SI §7). The results are visualized in Figure 3, which depicts boxplots for the fractions of time in confinement across all problem instances parameterized by c. We find that the dual-targeted AGE-ACT policy is able to reduce the confinement time quite systematically for every age group, relative to all other policies. Specifically, it results in the lowest confinement time for every age group in 70% of all problem instances when compared with NO-TARGET, in 60% of instances when compared with AGE, in 83% of instances when compared with ACT, and in 50% of instances when compared with all other policies. Moreover, the fraction of confinement time achieved by AGE-ACT is within 5% (in absolute terms) from the lowest confinement time achieved by any policy for every age group, in 76% of all instances; within 10% in 80% of the instances; and within 14% in all instances.
5/9 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Thus, even when the dual-targeted policy confines certain age groups more, it does not do so by much. These outcomes are quite unexpected as they are not something that the ROLD framework explicitly optimizes for, but rather a by-product of a dual-targeted confinement policy that minimizes the total loss objective (2). It is worth noting that although ROLD AGE-ACT reduces confinements for every age group compared to less targeted policies, it does not do so uniformly. Instead, it can lead to a larger discrepancy in the confinements faced by different age groups: those aged 10-59 are generally more confined than those aged 0-9 or 60+.
Figure 3. Average time in confinement for the ROLD policies with different targeting types. Each boxplot depicts the fraction of time the age group spends in confinement under the respective policy averaged over the activities relevant to that age-group, for different problem instances parameterized by the cost of death c.
The impact of limited disparity requirements. That targeted policies confine some age groups more than others could be perceived as disparate or unfair treatment, so it is important to quantify how an intervention’s effectiveness is impacted when requiring less differentiation across age groups. We embed a set of “limited disparity” constraints in ROLD that allow the activity levels of distinct age groups to differ by at most D in absolute terms, in each activity and at any point of time (SI §7). D = 0 corresponds to a strictly non-discriminatory policy, whereas a larger value of D allows more targeting, with D = 1 corresponding to a fully targeted policy. For every value of D, we record the total loss incurred by a ROLD policy with the limited disparity constraints, and calculate the increase in total loss relative to a fully targeted ROLD policy. We repeat the experiment for different problem instances parametrized by c, and Figure 4 depicts boxplots of all the relative increases in total loss, as a function of D. The results suggest that limited disparity requirements may be costly: on average, completely eliminating disparity in confinements would increase the total losses by EUR 1.2B (21.6%) and produce an additional 506 deaths (16.6%) and an extra EUR 0.5B of economic losses (18.9%) compared to a fully targeted policy. In certain instances, the increase in total loss could be as high as 63%. The high losses persist even when some limited discrepancy is allowed, dropping at an initially slow rate as D increases from 0 and eventually at a slightly faster rate as it approaches 1. This suggests that to fully leverage the benefits of targeting, a high level of disparity must be accepted, but reasonable trade-offs can be achieved with some intermediate disparity.
Figure 4. The impact of limited disparity requirements. The plot shows the relative (%) increase in total loss generated by a ROLD policy, compared to a fully targeted policy, as a function of the disparity parameter D that measures the maximum allowed difference in activity levels for distinct age groups. The experiments are run using several values of c, which are used to generate each of the boxplots. Eleven values of D are tested, ranging from 0 to 1.
6/9 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Discussion Our case study suggests that an optimized intervention targeting both age groups and activities carries significant promise for alleviating a pandemic’s health, economic and even psychological burden, but also points to certain challenges that require care in a real-world setting. Why consider optimized dual-targeted interventions? The first reason are the significantly better health and economic outcomes: for the same or a lower number of deaths, dual- targeted confinements can reduce economic losses more than any of the simpler interventions that uniformly confine age groups or activities. Furthermore, the super-additive gains imply that significant synergies can be generated through finer targeting, with the ability to target along activities improving the effectiveness of targeting along age groups, and vice-versa. The second reason is the intuitive nature of the optimized targeted confinement policy, which is consistent with a simple “bang-for-the-buck” rule: impose less confinement on group-activity pairs that generate a relatively high economic value prorated by (activity-specific) social contacts. This simple intuition combined with the reliance on just a few activity levels are appealing practical features, as they provide transparency into how targeted confinement decisions could be made. The simple “bang-for-the-buck” intuition is also related to the third benefit of dual targeting: the ability to impose less restrictive confinements across all age groups. As different age groups may be responsible for generating a larger economic value prorated by social contacts in distinct activities, dual-targeted confinements may enable all age groups to remain more active, resembling normal life more closely compared to less differentiated confinements. This could result in more socially acceptable restrictions, and a more appealing policy intervention overall. Lastly, we note that although dual targeting allows and can result in differences in confinements across age groups, such interventions are actually not far from many real-world policy implementations, which have been more or less explicit in their age-based discrimination. Dual targeting can arise implicitly in interventions that only seem to target activities. As an example, France implemented a population-wide 6 p.m. to 6 a.m. curfew (31), while maintaining school and work activities largely de-confined. This is effectively implementing restrictions similar to ROLD AGE-ACT: since a typical member of the 20-65 y.o. group is engaged in work until the start of the curfew, their leisure and other activities are implicitly limited; moreover, since most individuals aged above 65 are not in active employment, they are not that restricted in these last two activities. These examples show that some amount of targeting of activities and age groups is already in place and is perhaps unavoidable for effective pandemic management. Given this state of affairs, our framework highlights the significant benefits in explicitly and transparently modelling targeting and identifying the interventions that rigorously optimize overall societal welfare, given some allowable amount of differentiation across age groups. Challenges and limitations. An immediate practical challenge is data availability. Social contact matrices by age group and activity may be available from surveys on social behavior, which have been conducted in a number of countries; however, further data collection might be required to obtain these matrices for more refined population group or activity definitions. Similarly, economic data is reported by industry activities, but we are not aware of a dataset that splits economic value into separate (group, activity) contributions. Disparate data sources may be difficult to align: for example, social contact surveys and Google mobility reports use different activity categories, which requires non-trivial fitting (SI §4). Availability of data also constrains our model’s structure in several ways. In our model social contacts between age groups only depend on confinements in the same activity, since the available social contacts dataset (24) only reports interactions in the same activity. However, contacts occur as individuals are engaged in different activities (e.g., a professional in the services industry interacts during work with individuals who are engaged in leisure activities). A more refined contact mixing model that captures such interactions would be more appropriate for this study, provided that relevant data are available. Our economic model similarly ignores cross effects, such as young age groups engaged in school producing value in conjunction with educational staff engaged in work. Another challenge with targeted interventions is the perception that they lead to unfair outcomes, as certain age groups face more confinement than others. Such discrimination does arise in the optimized dual-targeted policies; our framework can partially address these concerns through explicit constraints that limit the disparities across groups. Our requirement that limited disparity hold for every time period and every activity is quite strict, and a looser requirement based, e.g., on time-average confinements could lead to smaller incremental losses. Alternatively, one can impose fairness requirements based on the intervention’s outcomes, e.g., requiring that the health or economic losses faced by different groups satisfy certain axiomatic fairness properties (32). Although we focus on confinement policies, a direction for future research is to investigate how these can be optimally combined with other types of targeted interventions. SI §7 reports experiments where we optimize a targeted policy based on confinements and randomized testing and quarantining. The framework is sufficiently flexible to accommodate interventions such as contact tracing and also vaccinations, although a careful implementation would require work beyond the scope of this article.
7/9 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Acknowledgments The authors thank Ciprian Badescu,˘ Florin Georgescu, Alexandru Hulea, Andrei Leica, and Vlad Paunescu˘ from Arnia Software for their efforts in implementing the framework and hosting it at http://insead.arnia.ro.
Authors’ contributions DFC, DAI and SIZ contributed equally to the research design, conceptualisation, modeling, analysis, and writing. SC and XSW contributed equally to the modeling, software implementation, data analysis and creating visualizations. DFC, DAI and SIZ verified all the data used in this study.
Declaration of interests All authors declare no competing interests.
References 1. Acemoglu, D., Chernozhukov, V., Werning, I. & Whinston, M. D. Optimal targeted lockdowns in a multi-group SIR model. Working Paper 27102, NBER (2020). DOI: 10.3386/w27102.
2. Matrajt, L., Eaton, J., Leung, T. & Brown, E. R. Vaccine optimization for COVID-19, who to vaccinate first? (2020).
3. Goldstein, J. R., Cassidy, T. & Wachter, K. W. Vaccinating the oldest against COVID-19 saves both the most lives and most years of life. Proc. Natl. Acad. Sci. 118, DOI: 10.1073/pnas.2026322118 (2021). https://www.pnas.org/content/118/ 11/e2026322118.full.pdf.
4. Bertsimas, D. et al. Optimizing vaccine allocation to combat the COVID-19 pandemic (2020).
5. Favero et al., C. A. Restarting the economy while saving lives under COVID-19 (2020).
6. Birge, J. R., Candogan, O. & Feng, Y. Reducing economic losses with targeted closures. Working Paper 2020-57, University of Chicago, Becker Friedman Institute for Economics (2020).
7. Chang, S. et al. Mobility network models of COVID-19 explain inequities and inform reopening. Nature 589, 1–6 (2020).
8. Evgeniou, T. et al. Epidemic models for personalised COVID-19 isolation and exit policies using clinical risk predictions, DOI: 10.1101/2020.04.29.20084707 (2020). https://www.medrxiv.org/content/early/2020/05/12/2020.04.29.20084707.full. pdf.
9. Giordano, G. et al. Modeling vaccination rollouts, SARS-CoV-2 variants and the requirement for non-pharmaceutical interventions in Italy. Nat. Medicine 27, 993–998, DOI: 10.1038/s41591-021-01334-5 (2021).
10. Bastani, H. et al. Deploying an artificial intelligence system for COVID-19 testing at the Greek border. Available at SSRN (2021).
11. Tiirinki, H. et al. COVID-19 pandemic in Finland–preliminary analysis on health system response and economic consequences. Heal. policy technology 9, 649–662 (2020).
12. Harrison, S. Coronavirus: Ireland’s restrictions eased for over 70s (2020). Access. 01/12/2021.
13. Magid, J. Cabinet approves removal of age restriction for those returning to work (2020). Accessed January 12, 2021.
14. Foy, H. Russian businesses prepare for fresh lockdowns as Covid-19 cases soar (2020). Accessed January 12, 2021.
15. Reuters Staff. Bosnian region eases lockdown on seniors, children after court ruling (2020). Accessed January 12, 2021.
16. Kanbur, N. & Ankgül, S. Quaranteenagers: A single country pandemic curfew targeting adolescents in Turkey. J. Adolesc. Heal. 67 (2020).
17. Kucharski, A. J. et al. Effectiveness of isolation, testing, contact tracing, and physical distancing on reducing transmission of sars-cov-2 in different settings: a mathematical modelling study. The Lancet Infect. Dis. 20, 1151–1160, DOI: 10.1016/S1473-3099(20)30457-6 (2020).
8/9 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
18. Prem, K. et al. The effect of control strategies to reduce social mixing on outcomes of the COVID-19 epidemic in Wuhan, China: a modelling study. The Lancet Public Heal. 5, e261–e270, DOI: 10.1016/S2468-2667(20)30073-6 (2020).
19. Di Domenico, L., Pullano, G., Sabbatini, C. E., Boëlle, P.-Y. & Colizza, V. Impact of lockdown on COVID-19 epidemic in Île-de-France and possible exit strategies. BMC medicine 18, 1–13 (2020). 20. Aguilera, J. Some supermarkets are launching senior-only hours during the coronavirus pandemic. Not all retailers think that’s a good idea (2020). Accessed January 12, 2021.
21. Anderson, R. M. & May, R. M. Infectious Diseases of Humans: Dynamics and Control (Oxford University Press, 1992). 22. Salje, H. et al. Estimating the burden of SARS-CoV-2 in France. Science 38, DOI: 10.1126/science.abc3517 (2020). 23. Wille, L. et al. SOCRATES: an online tool leveraging a social contact data sharing initiative to assess mitigation strategies for COVID-19. BMC Res. Notes 13 (2020).
24. Béraud, G. et al. The French connection: The first large population-based contact survey in France relevant for the spread of infectious diseases. PLOS ONE 10, 1–22, DOI: 10.1371/journal.pone.0133203 (2015). 25. French Government. Données hospitalières relatives à l’épidémie de COVID-19. https://www.data.gouv.fr/fr/datasets/ donnees-hospitalieres-relatives-a-lepidemie-de-covid-19/ (2020). Accessed October 21, 2020.
26. Google. COVID-19 community mobility report. https://www.google.com/covid19/mobility/ (2020). Accessed October 21, 2020.
27. Borrelli, F., Bemporad, A. & Morari, M. Predictive control for linear and hybrid systems (Cambridge University Press, 2017).
28. Duque, D. et al. Timing social distancing to avert unmanageable COVID-19 hospital surges. Proc. Natl. Acad. Sci. 117, 19873–19878 (2020).
29. Lehot, M. & Borgne, B. L. Covid-19 : ces chiffres qui montrent que Paris a dépassé le seuil d’alerte maximale depuis le 25 septembre (2020). Accessed October 5, 2020.
30. Alvarez, F. E., Argente, D. & Lippi, F. A simple planning problem for COVID-19 lockdown. Working Paper 26981, National Bureau of Economic Research (2020). DOI: 10.3386/w26981.
31. Reuters. France goes under nationwide 6pm curfew as Covid-19 death toll surpasses 70,000 (2021). Accessed January 12, 2021.
32. Young, H. P. Equity (Princeton University Press, 1994).
9/9 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Supplementary Information for
Quantifying The Benefits of Targeting for Pandemic Response
Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis
Emails of authors: [email protected], fl[email protected], [email protected], [email protected], [email protected]
This PDF file includes: Supplementary text Figs. S1 to S10 Tables S1 to S8 SI References
Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis 1 of 35 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Supporting Information Text 1. Some Notation We denote scalars by lower-case letters, as in v, and vectors by bold letters, as in v. We use square brackets to denote the concatenation into vectors: v := [v0,v1]. For a time series of vectors v1,...,vn, we use the notation vi:j := [vi,...,vj ] to denote the concatenation of vectors vi through vj . Lastly, we use v€ to refer to the transpose of v.
2. Model and Optimization Problem In this section, we propose a modified version of the discretized SEIR (Susceptible - Exposed - Infectious - Recovered) population-based epidemiological model with multiple population groups, which interact with each other. In our model, the SEIR dynamics are controllable via (i) choosing confinement patterns for di erent groups and (ii) choosing how to apportion testing among the groups. Then, we build a economic model on top of the SEIR system, which factors in the cost of lost output, together with the cost of deaths. We conclude by formulating the overall optimization problem.
A. Epidemiological Model. Time. Time is discrete, indexed by t =0, 1,...,T, and measured in days. We assume that no infections are possible beyond day T .
Age Groups. We split the population into age groups. We use to denote the collection of considered age groups. In our G case study for Île-de-France we use nine age groups, split in 10-year buckets, where the youngest group are the 0-9 year olds (y.o.), and the oldest group are the 80+ year olds. Since individuals in separate groups have social contacts with each other, the dynamics of each group depend on the number of infections in other groups.
Compartmental Model and States. At any given time, the population of an age group is divided into susceptible, exposed, infectious, recovered, and deceased as in a classical SEIR model. Figure S1 represents the compartmental model and the SEIR transitions for a specific group g. All the states represent the number of individuals in a compartment of the model in the beginning of the time period. State Sg(t) is the number of individuals in group g that are susceptible to be infected at time t. State Eg(t) is the number of individuals in group g that have been exposed to the SARS-CoV-2 virus at time t, but are not yet infectious. Ig(t) is the number of individuals in group g that are infectious at time t, and have not yet been tested, or recovered, or transferred to the hospital or the ICU. Susceptible individuals get infected and transition to the exposed state at a rate determined by the number of social contacts as well as transmission rate —(t). Exposed individuals transition to the infectious state at a rate ‡. q Ig (t) is the number of individuals in group g who have been confirmed to be infectious through viral testing at time t, and are thus quarantined. We subdivide Iq(t) into Iq (t) for j a, ps, ms, ss to model di erent degrees of severity of symptoms: g j,g œ{ } infectious individuals can be asymptomatic, paucisymptomatic, have mild symptoms, or have severe symptoms. We assume that an infectious individual in group g will exhibit symptoms of degree j with probability pj,g. Infectious individuals transition out of the infectious state at a rate µ. Individuals with severe symptoms need hospitalization. Hg(t) is the number of individuals in group g that are in the general hospital wards at time t, and ICUg(t) is the number of individuals in group g that are in the intensive care unit at time t.We H ICU assume that an infectious individual will need to be hospitalized in the general wards (go to ICU) with probability pg (pg ). H ICU 1 We have pss,g = pg + pg . On average, patients who are treated in the general hospital wards spend ⁄H≠ days in the hospital; 1 patients who are treated in the ICU spend ⁄ICU≠ days in the ICU. Infectious individuals with severe symptoms may decease. Dg(t) is the number of individuals in group g that have died from COVID-19 up until time t. We assume that an infectious individual with severe symptoms in group g will decease with probability pD (and recover with probability pR =1 pD). g g ≠ g Infectious individuals usually recover. Rg(t) is the number of individuals in group g who have recovered until time t, but q have not been confirmed to have recovered. Rg(t) is the number of individuals in group g who have recovered until time t, and have been confirmed to have recovered and to have had the virus either through testing, or because they recovered during their stay in the hospital or the ICU. We keep track of the total number of individuals in group g who are not confirmed infectious, are not confirmed recovered, are not in the hospital or the ICU, and have not died: Ng(t):=Sg(t)+Eg(t)+Ig(t)+Rg(t).Welet
X(t)= Sg(t),Ig(t),...,Dg(t) g œG denote the full state of the system at time t =0, 1,...,T# 1,T, i.e., the number$ of individuals in each of the states, across ≠ groups. We denote the number of compartments by . Then the dimension of X(t) is 1. |X | |G||X | ◊ We assume that Dg(0) = 0 for all g . œG
2 of 35 Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Fig. S1. Compartmental SEIR model for a specific group g
q q Ij,g ICUg Rg
q Sg Eg Ig Iss,g Hg Dg
SEIR transition Rg Transition due to viral testing
Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis 3 of 35 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
A.1. Controls/Decision Variables. We allow for three levers for the policy maker’s response: targeted testing, targeted confinements, and targeted patient turn-away decisions.
Confinements and social mixing model. We refer to the activity level decision for a group g as the vector
home work school transport leisure other 6 ¸g(t):= ¸ (t),¸ (t),¸ (t),¸ (t),¸ (t),¸ (t) [0, 1] g g g g g g œ that captures the level of each activity# that an individual in group g is allowed to perform$ in period t, in relation to the normal activity level under normal course (no confinement), as a result of imposed confinement restrictions. We denote the set of considered activities by := home, work, school, transport, leisure, other . A { } work Normal course, i.e., no confinement, is modeled using ¸g = 1 =[1, 1, 1, 1, 1, 1]. As an example, ¸g =0.7 can be interpreted as a restriction for individuals in group g to spend at the workplace only 70% of the baseline time they spend at the workplace. We provide more details about how the baseline is defined for our numerical experiments and calibration in Section 4 B. When unambiguous, we use the terms activity levels and confinements to refer to the same underlying decision of ¸g(t). The confinement decision variables for the policy maker are the activity levels of each group, in each time period: ¸g(t) g ,t=0:T 1. { } œG ≠ In our proposed implementation for Île-de-France we allow the confinement decisions to update every two weeks. Confinements a ect the amount of social contacts between individuals of di erent groups. We build a model to reflect this dependence which we describe below. For activity a and groups g, h, we denote Ca the number of contacts in a day between an individual in group g and œA g,h individuals in group h in activity a under normal course (no confinement). We retrieve the contact matrices per activity Ca,a from the social contact data tool by (1), using the survey on social contacts in France by (2). œA Let cg,h(¸g, ¸h) denote the mean number of contacts per day between an individual in group g and individuals in group h, when groups g, h are confined according to confinement patterns ¸g, ¸h, respectively. We use the following multiplicative extrapolation of social mixing matrices, with parameter –1,–2 R: œ
a a –1 a –2 cg,h(¸g, ¸h)= C ¸ (¸ ) . [1] g,h · g · h a ÿœA ! " We remark that this parametrization is similar to a Cobb-Douglas production function (3), using the confinement patterns as inputs, and the number of social contacts as output. In Section 4 B, we describe our procedure for fitting the social mixing parameters from data on health outcomes and mobility data.
Testing. We model viral diagnostic tests. Individuals who are found to be infected through a diagnostic test are quarantined. We assume that there is access to perfect tests (i.e., the sensitivity and specificity are 100%). We assume that a viral test yields a positive result when conducted on an infectious individual (compartment Ig), and a negative result when conducted on an individual who is in a non-infectious state. Vtest We use Ng (t) to denote the number of viral tests allocated to group g in period t. The testing decisions for the policy Vtest maker are then Ng (t) g ,t=0:T 1.We assume that, given a testing allocation decision across groups, the individuals { } œG ≠ to receive the allocated tests in group g are randomly chosen among individuals in Ng(t). For example, the outflow from Vtest performing Ng (t) viral tests in group g, which happens out of compartment Ig(t), is equal to
Ig(t) Vtest Ng (t). [2] Ng(t) ·
In our proposed implementation for Île-de-France we allow the testing decisions to change weekly. Figure S1 represents the compartmental model and the flows resulting from testing from one compartment to the other for a specific group g.
Turning patients away. When the patient inflow into the hospital or the ICU exceeds the remaining number of available beds, then the decision maker needs to decide how to prioritize the admission of patients to the available beds. Our general formulation allows the decision maker to optimize the number of patients from each group turned away from the general hospital wards and from the ICU, at each time period. We denote the number of patients from age group g who need to be H admitted to the general hospital wards but are turned away in period t by Bg (t). We denote the number of patients from age ICU group g who need to be admitted to the ICU but are turned away in period t by Bg (t). The turn-away decision variables for H ICU the policy maker are then Bg (t),Bg (t) g ,t=0:T 1. We assume that all patients who are turned away decease immediately. { } œG ≠ In our experiments we use a proportional turn-away rule, which is defined in Section A.3.
Let Vtest H ICU u(t)= ¸g(t),N (t),B (t),B (t) g g g g œG denote the vector of all the decisions/controls at# time t =0, 1,...,T 1, i.e., the$ confinement and viral test decisions for all ≠ the groups, as well as the decisions on turning patients away. We denote the number of di erent decisions for a given group at agiventimeby . Then the dimension of u(t) is 1. |U| |G||U| ◊
4 of 35 Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
A.2. Dynamics of the Controlled SEIR Model. Having defined the states and controls, we write down a set of discrete time dynamics for the controlled SEIR model. We use notation X(t) to denote X(t +1) X(t). For all groups g , we write ≠ œG
Vtest Ig(t) H ICU Ng(t)= Ng (t) µ (pg + pg ) Ig(t) [3] ≠ · Ng(t) ≠ · · infected going to ICU, H
¸ ˚˙ Ih(˝t) Sg(t)= —(t)Sg(t) cg,h(¸g(t),¸h(t)) q [4] ≠ · Nh(t)+R (t) Ah h B ÿœG Ih(t) Eg(t)=—(t)Sg(t) cg,h(¸g(t),¸h(t)) q ‡Eg(t) [5] · Nh(t)+R (t) ≠ Ah h B ÿœG H ICU H ICU Vtest Ig(t) Ig(t)=‡Eg(t) µ (pg + pg ) Ig(t) µ (1 pg pg ) Ig(t) Ng (t) [6] ≠ · · ≠ · ≠ ≠ · ≠ · Ng(t) go to ICU, H recover q outflow to I ,g from positive V tests · H¸ ICU ˚˙ ˝ ¸ ˚˙ ˝ Rg(t)=µ (1 pg pg ) Ig(t) ¸ ˚˙ ˝ [7] · ≠ ≠ · q Vtest Ig(t) q Ij,g(t)= pj,g Ng (t) µ Ij,g(t), j a, ps, ms [8] · · Ng(t) ≠ · ’ œ{ } recover inflow from Ig from positive V tests
q Vtest Ig(t) ¸ ˚˙ ˝ q Iss,g(t)= ¸pss,g Ng ˚˙ (t) ˝ µ Iss,g(t) [9] · · Ng(t) ≠ · go to ICU, H, R, or D inflow from Iss,g from positive V tests
q q H R ¸ ˚˙ICU ˝ R Rg(t)=µ ¸ I˚˙ (t)+⁄g ˝ pg Hg(t)+⁄g pg ICUg(t) [10] · j,g · · · · j a,ps,ms œ{ ÿ } recovery of known infected cases H ¸ H H ˚˙ pg q H ˝ Hg(t)= ⁄g Hg(t)+µ pg Ig(t)+ H ICU Iss,g(t) Bg (t) [11] ≠ · · pg + pg · ≠ 3 4 ICU ICU ICU pg q ICU ICUg(t)= ⁄g ICUg(t)+µ pg Ig(t)+ H ICU Iss,g(t) Bg (t) [12] ≠ · · pg + pg · ≠ 3 4 H D ICU D H ICU Dg(t)=⁄ p Hg(t)+⁄ p ICUg(t)+B (t)+B (t). [13] g · g · g · g · g g
In Eq. (7) and Eq. (10), note that we do not have terms for the population turned away from hospital/ICU which may eventually recover. Instead, we assume the turned away patients will go into the deceased state. In Eq. (13), we are assuming that if a patient is turned away from the ICU, they transition into deceased, instead of being allocated a hospital bed if one is available.
We can summarize the SEIR dynamics with
X(t +1)=X(t)+Ft (X(t), u(t)) , [14]
where the function Ft captures the dynamics in Eq. (3)-Eq. (13).
We now provide justification for how we account for social contacts and, in particular, for the expressions in Eq. (4) q and Eq. (5). We assume that individuals in Rg can interact with members of Sg,Eg,Ig and Rg but cannot infect.
Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis 5 of 35 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
Fix a person i in age group g ,instateSg. Then we can write œG
Pr i got infected through socializing with age group h [15] { } Ah B €œG
= Pr i got infected through socializing with individuals in Sh,Eh,Ih or Rh [16] { } Ah B €œG =1 Pr( i did not get infected through socializing with individuals in any Sh,Eh,Ih or Rh for any h ) [17] ≠ { œG} =1 Pr( i did not get infected through socializing with individuals in Sh,Eh,Ih or Rh ) [18] ≠ { } h ŸœG Nh(t) q cg,h Ih(t) Nh(t)+R (t) =1 1 —(t) h [19] ≠ ≠ Nh(t) h 3 4 ŸœG Nh(t) Ih(t) 1 1 —(t) q cg,h [20] ¥ ≠ ≠ · Nh(t)+R (t) · · Nh(t) h 3 h 4 ŸœG Ih(t) —(t) cg,h q , [21] ¥ Nh(t)+R (t) h h ÿœG where in Eq. (19) we use the following reasoning. Having fixed person i in age group g, the number of her contacts in age group h is given by cg,h = cg,h(¸g(t),¸h(t)). The number of contacts cg,h reflects the number of contacts of individual i from group h, assuming all individuals in age group h are put under the same confinement pattern. However, recall that individuals q q in Rh(t) cannot infect individuals in Sh,Eh,Ih or Rh. Therefore, we scale cg,h by the ratio of (S, E, I, R, R ) individuals in Nh(t) age group h that are in states (S, E, I, R),i.e., N t Rq t , to account for the role of the relative size of groups of individuals h( )+ h( ) in di erent confinement patterns within an age group. Finally, getting infected as the result of any contact within group h is considered to be an independent event. By taking the expectation of random variable
i got infected through socializing , { } i S ÿœ g we retrieve the expressions in Eq. (4) and Eq. (5).
A.3. Resources and Constraints. We use KH(t) (KICU(t)) to denote the capacity of beds in the general hospital wards (in the ICU) on day t. When the patient inflow into the hospital or the ICU exceeds the remaining number of available beds, then the policy maker needs to decide how many patients to turn away from each group. Although our framework allows to optimize over the decisions of turning away patients, we mainly look at a specific turn-away rule: let in patients from each age group proportionally to the arrivals from that age group, up to capacity, and turn away the remaining patients. This specific turn-away rule can be expressed by setting
H H pg q µ pg Ig(t)+ H ICU Iss,g(t) H · pg +pg · Bg (t)= pH 1 H h q 2 µ p Ih t I t h h ( )+ pH +pICU ss,h( ) · h h · 1 2 q proportion of inflow into H from group g + ¸ ˚˙ ˝ H H ph q H H Q µ p Ih(t)+ I (t) K (t) 1 ⁄ Hh(t) R [22] · h · pH pICU · ss,h ≠ ≠ ≠ g h + h A B c h 3 4 h d cÿ ÿ ! " d c total inflow into H available beds in H d a ICU b ICU pg q µ¸ pg Ig(t)+ H ˚˙ ICU Iss,g(t) ˝ ¸ ˚˙ ˝ ICU · pg +pg · Bg (t)= pICU 1 ICU h q 2 µ p Ih t I t h h ( )+ pH +pICU ss,h( ) · h h · 1 2 + q ICU ICU ph q ICU ICU µ p Ih(t)+ I (t) K (t) 1 ⁄ ICUh(t) [23] · h · pH + pICU · ss,h ≠ ≠ ≠ g A h 3 h h 4 A h BB ÿ ÿ ! " We further assume a given capacity for viral tests each day, which we denote by KVtest(t) on day t. We assume that viral tests used to test individuals with severe symptoms that enter the ICU or hospital, as well as viral tests that test hospitalized
6 of 35 Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
individuals to confirm they have recovered, come from a di erent pool of tests and do not consume the capacity for viral testing in the non-hospitalized population. We can now write the following constraints for the optimization problem:
H (“Hospital capacity”) Hg(t) K (t), t [24] Æ ’ g ÿ H H H pg q (“Hospital turn away”) Bg (t) µ pg Ig(t)+ H ICU Iss,g(t) 0, g, t [25] ≠ · pg + pg · Æ ’ 3 4 ICU (“ICU capacity”) ICUg(t) K (t), t [26] Æ ’ g ÿ ICU ICU ICU pg q (“ICU turn away”) Bg (t) µ pg Ig(t)+ H ICU Iss,g(t) 0, g, t [27] ≠ · pg + pg · Æ ’ 3 4 (“Viral testing capacity”) N Vtest(t) KVtest(t), t [28] g Æ ’ g ÿ (“Fractional lockdown”) ¸a(t) 1, g, a, t [29] g Æ ’ (“Non-negative decisions”) N Vtest(t),¸a(t) 0,BH (t) 0,BICU(t) 0 g, a, t [30] g g Ø g Ø g Ø ’ We denote by (X(t)) the feasible set described by Eq. (24)-Eq.(30) for the vector of decisions u(t) at time t. C B. Economic Model and Optimization Objective. Here, we describe in more detail the objective we use in our optimization problem. This objective seeks to combine two di erent losses due to the pandemic: (i) the measurable economic loss stemming from quarantining, confinements and wages lost due to death and illness as well as (ii) the mortality from the pandemic. We first describe how we account for economic loss in our model.
B.1. Economic Loss. These losses come from three separate sources:
E ect of quarantining. We capture the e ect of quarantine and hospitalization by assuming that if at some time pe- q riod an individual in group g is in one of the SEIR chambers Ij,g, j a, ps, ms, ss ,Hg, ICUg, then they generate no q ’ œ{ } economic value. At the same time, we assume that individuals in Rg generate economic value as they would under no confinement.
E ect of confinement. To account for confinement in the non-quarantined population, we make the economic value per day generated by an individual in group g in the remaining (non-quarantined) SEIR chambers explicitly depend on the enforced confinement in the population. Recall that for a group g, the activity levels ¸g specify the level of each activity allowed for that group as compared to normal course, and ¸ = ¸g g . We denote the economic value generated by a member of g per day by { } œG vg(¸). We remark that vg(1) corresponds to the economic value generated by an individual under normal circumstances. The vg(¸) specific to a group can be of two types: (a) wages from employment and (b) future wages from employment due to schooling. Naturally, depending on the age group, both, one, or neither of these will actually contribute to economic value. Distinguishing whether the specific group is comprised of school age, employable or retired population, we define
schooling vg (¸) if g =0-9y.o. employment schooling vg (¸)+vg (¸) if g =10-19y.o. vg(¸):=Y [31] vemployment(¸) if g =20-29,30-39,40-49,50-59,60-69y.o. _ g ]0 otherwise.
_ schooling employment We break down the definitions[ of vg (¸) and vg (¸) below:
employment • Value from employment vg (¸). The value generated from employment is a function of the confinement level in the work activity, but also of the confinement levels in leisure, transport,aswellasother activities. As an example, we expect the economic value of those employed in restaurants, retail stores, etc. to depend on foot tra c levels, which in turn are driven by the confinement levels in leisure, transport and other activities across all groups. employment Our model for employment value is a linear parametrization of these confinement decisions; specifically, vg (¸) is linear in ¸work and the average of ¸transport, ¸leisure and ¸other across these three activities and all groups g : g œG
transport leisure other employment work work other activities 1 ¸h + ¸h + ¸h fixed v (¸):=wg ‹ ¸ + ‹ + ‹ . [32] g · · g · 3 A A |G| h B B ÿœG Additionally, ‹work,‹other activities and ‹fixed are activity level sensitivity parameters such that ‹work 1+‹other activities 1+ fixed · · ‹ =1; under fully open activity, they induce a multiplier of 1 in Eq. (32). Then, wg measures the overall employment employment value of a typical member of group g under no confinement, and is equal to vg (1).
Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis 7 of 35 medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
We estimate the coe cients of this model from data, as we describe in detail in Section 4.
schooling • Value from scho oling vg (¸). A day of schooling for the individuals in relevant groups results in economic value, equal to a day of wages that members of these groups would gain in the future. We use the salary of the 20-29 year-old group, multiplied by a factor, and we discount for a number of years corresponding to the di erence between the midpoint of the age group and the beginning of the 20-29 year-old group. For instance, for the 0-9 year olds, we would discount over 15 years, and for 10-19 year olds we discount over five years. That is, the discounting factor we apply is 15 (1 + r)≠ if g = 0-9 y.o. 5 ”g := (1 + r)≠ if g = 10-19 y.o., Y ]0 otherwise, where r is the discount rate. We further multiply[ the wage by fg, which is the fraction of group g that is in school. Lastly, we also use a multiplicative factor ◊. We use this for sensitivity analysis: ◊ reflects that an additional day of schooling may have a multiplier e ect in future wages, as well as the fact that schooling can be continued online during lockdowns. We provide ranges for ◊ in Section 6. Thus, the definition for value of school days becomes
schooling employment school v (¸):=◊ fg ”g v (1) ¸ . [33] g · · · 20-29 y.o · g
E ect of lost wages due to death. We account for a deceased individual’s loss of wages which they would have earned from their current age until retirement age, given the prevailing wage curve under status quo conditions given by vg(1) g . { } œG For group g, we set the current age to the mid-point of the age group. We discount the resulting cash flows by an annualized life interest rate. We denote the resulting lost wages amount by vg . For instance, for someone in age group 30-39 y.o., we calculate this cash flow by†
69 1 vlife 30 39 y.o. := · 35 ≠ (1 + r) ≠ · ÿ=35 1 employment 1 employment 35 · 39 v30 39 y.o. (1)+ 40 · 49 v40 49 y.o. (1) · { Æ Æ }· ≠ { Æ Æ }· ≠ 1 employment 1 employment + 50 · 59 v50 59 y.o. (1)+ 60 · 69 v60 69 y.o. (1) . [34] ! { Æ Æ }· ≠ { Æ Æ }· ≠ We are now ready to define the economic loss component of the objective. We first define a quantity V "which represents the economic value that would be generated in all groups g under a “no pandemic” scenario. More precisely, to calculate V we œG assume that at time t =0all the infected or exposed population is instantaneously healed and able to generate full economic value vg(1). Thus,
T 1 ≠ q q V := vg(1) Ng(0) + I (0) + Hg(0) + ICUg(0) + R (0) . [35] · Q j,g g R t=0 g j a,ps,ms,ss ÿ ÿœG œ{ ÿ } a b We note that this term is a constant and does not depend on the policy followed by the policy maker. The economic loss is then defined as the di erence between V and the (potentially negative) value generated during the pandemic, and which depends on the policy followed by the policy maker:
T 1 ≠ q life Economic Loss(u0:T 1):=V vg (¸(t)) (Sg(t)+Eg(t)+Ig(t)+Rg(t)) + vg(1) Rg(t) + vg Dg(T ), [36] ≠ ≠ · · · t=0 g g ÿ ÿœG ! " ÿœG B.2. Total Deaths. We associate each death with a non-pecuniary cost parametrized by ‰. Similarly to (4), our approach is to characterize the frontier between deaths and economic losses. For any cost of death ‰, maximizing the objective function for that ‰ will yield a particular point on the frontier. One can trace the frontier by varying ‰. We focus on characterizing the frontier, rather than selecting an optimal point along the frontier. Selecting an optimum would entail determining a choice for the value of life, which is not desirable, as there is disagreement about the right value. We define
Total Deaths(u0:T 1):= Dg(T ). [37] ≠ g ÿœG This is due to the fact that a small fraction of the members of the 10-19 year old group are already in workforce. We do not count the value of lost schooling for them. †Note that we are counting the entire value for the 60-69 y.o. age group; this is due to the fact that, as we explain in Section 4, this value has already incorporated that only a fraction of the population in the 60-69 y.o. age group have work-eligible ages (60-64 y.o.), while the rest are retired.
8 of 35 Sergio Camelo, Dragos F. Ciocan, Dan A. Iancu, Xavier S. Warnes, and Spyros I. Zoumpoulis medRxiv preprint doi: https://doi.org/10.1101/2021.03.23.21254155; this version posted June 18, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC 4.0 International license .
C. Optimization Problem. We can now formulate the overall optimization problem we are solving. Our objective captures the trade-o between the cost of the pandemic in terms of lost lives, and the cost in terms of foregone economic output (the economic loss). In addition, we require that the state trajectory follow the SEIR dynamics, and that the controls and state trajectory respect the capacity and feasibility constraints discussed above. We can then formulate this problem as
min Economic Loss(u0:T 1)+‰ Total Deaths(u0:T 1) [38] u0:T 1 ≠ · ≠ ≠ s.t. X(t)+Ft (X(t), u(t)) , 0 t T 1 [39] ’ Æ Æ ≠ u(t) (X(t)), 0 t T 1. [40] œC ’ Æ Æ ≠
3. ROLD: Re-Optimization with Linearized Dynamics In this section, we describe our approach to solving problem Eq. (38)-Eq. (40) via a linearization heuristic. Before delving into this, we make a few comments on why is it di cult to solve this problem to optimality. Observe that a key dynamic in any SEIR-type model is the rate of new infections that multiplies the current susceptible population in a given group with the infected population in another group and which already introduces non-linearity in the state trajectory. This can be illustrated with our dynamic for the evolution of the susceptible population in group g,Eq.(4), which we replicate here:
Ih(t) Sg(t)= —(t) Sg(t) cg,h(¸g(t),¸h(t)) q . ≠ · · · Nh(t)+R (t) Ah h B ÿœG Vtest It can be easily seen that expanding out Sg(t) produces a nested fraction of polynomials in all the past decisions ¸(·),N (·) for 0 · t 1‡. This function has no identifiable structure that would make the resulting optimization problem tractable via Æ Æ ≠ convex optimization techniques. Similarly, the objective also involves products of states and controls, and su ers from the same non-linearity. Moreover, this issue persists even if one focuses on holding testing fixed and optimizing over the confinement decisions, or vice-versa. With this in mind we focus on developing heuristics that can tractably yield good policies, and propose an algorithm, which we call Re-optimization with Linearized Dynamics, or ROLD, that builds a control policy by incrementally solving linear approximations of the true SEIR system.
A. Linearization. The key idea here is to solve the problem in a shrinking-horizon fashion, where at each time step k =0,...,T we linearize the system dynamics and objective (over the remaining horizon), determine optimal confinement and testing decisions for all k, . . . , T, and only implement the decisions for the current time-step. We first describe the linearization of the dynamics. As a slight abuse of notation which helps readability, we index time with subscripts, i.e. we write Xt := X(t) and similarly, ut := u(t). Note that we can write the true evolution of our dynamical system as: